@johpaz/hive-sdk 0.1.4 → 0.1.6

package/packages/core/src/agent/skill-selector.ts

@@ -0,0 +1,374 @@
+/**
+ * HiveDB-based Dynamic Skill Selector Module
+ *
+ * Context Compiler Level 4 - Intelligent Skill Selection
+ *
+ * This module uses HiveDB BM25 scoring (Spanish stemming + accent folding)
+ * to select the most relevant skills based on the user message, similar to
+ * tool selection.
+ *
+ * DESIGN DECISIONS:
+ *
+ * 1. Reads from skills table in database (not hardcoded catalog)
+ * 2. Maximum skills per turn for balanced context injection
+ * 3. Explicit triggers checked first; search is the semantic fallback
+ * 4. Relative score cutoff (fraction of the top hit), never absolute
+ * 5. Returns skill content for injection into system prompt
+ */
+
+import { col } from "../storage/hive"
+import type { SkillDoc } from "../storage/collections"
+import { logger } from "../utils/logger"
+import { isMinimalSkill } from "./minimal-loadout"
+import { isCalendarOperation } from "./routing-intent"
+import {
+    searchCapabilities,
+    applyRelativeCutoff,
+    replaceCapabilityDocs,
+    type CapabilityDoc,
+} from "./capability-search"
+
+const log = logger.child("skill-selector")
+
+
+
+// ─── Types ───────────────────────────────────────────────────────────────────────
+
+export interface SkillDescriptor {
+    id: string
+    name: string
+    description: string
+    category: string
+    tools: string
+    triggers: string
+    preferred_agents: string
+    body: string
+    version: string
+    version_num: number
+    active: boolean
+}
+
+/**
+ * Every bundled SKILL.md self-titles right after its frontmatter (e.g.
+ * "# Canvas Report Skill\n\n## Cuándo se Activa...") for standalone
+ * readability — but every system-prompt injection site wraps the body with
+ * its own `## <skill.name>` header, so the leading H1 becomes a *higher*
+ * heading level nested under a lower one (malformed hierarchy) and repeats
+ * the same title twice, for free tokens on every turn. Strip it here, once,
+ * so the 29 source files keep their own title for anyone reading them
+ * standalone, while every injected copy stays clean.
+ */
+function stripRedundantTitle(body: string): string {
+    // Bodies carry a leading blank line from the frontmatter's closing "---",
+    // so the H1 isn't at position 0 — allow leading whitespace before it.
+    return body.replace(/^\s*# .+\n+/, "")
+}
+
+function toSkillDescriptor(doc: SkillDoc): SkillDescriptor {
+    return {
+        id: doc.id, name: doc.name, description: doc.description ?? "", category: doc.category,
+        tools: doc.tools, triggers: doc.triggers, preferred_agents: doc.preferred_agents,
+        body: stripRedundantTitle(doc.body), version: doc.version, version_num: doc.version_num, active: doc.active,
+    }
+}
+
+export interface SelectedSkill {
+    id: string
+    name: string
+    score: number
+    category: string
+    description: string
+    body: string
+}
+
+export interface SkillSelectorResult {
+    skills: SkillDescriptor[]
+    selected: SelectedSkill[]
+    reasoning: string
+    timingMs: number
+}
+
+// ─── Configuration ─────────────────────────────────────────────────────────
+
+/** Maximum skills to return per message */
+const MAX_SKILLS_PER_TURN = 4  // Increased from 2 to allow more skills
+
+/**
+ * Relative relevance cutoff: keep a hit only if it scores at least this
+ * fraction of the top hit. HiveDB BM25 scores are positive (higher = better)
+ * but corpus-dependent, so absolute thresholds don't transfer — a ratio does.
+ */
+const RELEVANCE_RATIO = 0.3
+
+/** Stopwords to filter out before BM25 query construction */
+const STOPWORDS = new Set([
+    "que", "con", "para", "por", "una", "uno", "los", "las", "del",
+    "como", "esta", "esto", "ese", "eso", "the", "and", "for",
+    "with", "this", "that", "have", "will", "also", "de", "en",
+    "el", "la", "se", "su", "sus", "al", "es", "son", "pero",
+    "más", "mas", "ya", "yo", "tu", "te", "ti", "mi", "me",
+    "hola", "hi", "hello", "hey", "gracias", "thank", "please",
+    "ok", "okay", "yes", "si", "no", "bien", "good", "great",
+    "puedes", "necesito", "quiero", "podés", "necesitás", "querés",
+])
+
+/** Conversational patterns that should return empty skill list */
+const CONVERSATIONAL_PATTERNS = [
+    /^(hola|hi|hello|hey|buenos? días?|buenas? noches?|qué tal|howdy)/i,
+    /^(gracias|thank you|thanks|muchas gracias|muchas thanks)/i,
+    /^(cómo estás?|how are you?|qué流水|you doing|qué cuentas)/i,
+    /^(sí|yes|ok|okay|de acuerdo|perfecto|claro|por supuesto)/i,
+    /^(adiós|bye|nos vemos|see you|later|chau)/i,
+    /^(entiendo|understand|i see|ya veo|got it)/i,
+    /^(bien|good|great|excelente|awesome|perfect)/i,
+    /^(?:\?|¿)$/,  // Just a question mark
+]
+
+// ─── Helper Functions ───────────────────────────────────────────────────────
+
+/**
+ * Check if message is purely conversational (no skills needed)
+ */
+function isConversational(message: string): boolean {
+    const trimmed = message.trim()
+
+    // Empty or very short messages
+    if (trimmed.length < 2) return true
+
+    // Check conversational patterns
+    for (const pattern of CONVERSATIONAL_PATTERNS) {
+        if (pattern.test(trimmed)) {
+            log.debug(`[skill-selector] Message matched conversational pattern: ${pattern}`)
+            return true
+        }
+    }
+
+    // Check if all words are stopwords (likely conversational)
+    const words = trimmed.toLowerCase().split(/\s+/)
+    const meaningfulWords = words.filter(w => w.length > 2 && !STOPWORDS.has(w))
+    if (meaningfulWords.length === 0) {
+        log.debug(`[skill-selector] All words are stopwords - conversational`)
+        return true
+    }
+
+    return false
+}
+
+/**
+ * Check if message matches explicit triggers from a skill
+ */
+function matchTriggers(message: string, triggersJson: string | null): boolean {
+    if (!triggersJson) return false
+
+    try {
+        // Triggers are stored as comma-separated string in DB (e.g., "trigger1,trigger2")
+        const triggers: string[] = triggersJson.split(",").map(t => t.trim()).filter(t => t.length > 0)
+        if (triggers.length === 0) return false
+
+        const lowerMessage = message.toLowerCase()
+        return triggers.some(trigger =>
+            lowerMessage.includes(trigger.toLowerCase())
+        )
+    } catch (err) {
+        log.warn(`[skill-selector] Failed to parse triggers: ${(err as Error).message}`)
+        return false
+    }
+}
+
+// ─── Main Selection Function ─────────────────────────────────────────────────
+
+/**
+ * Select skills for a given user message using hybrid matching:
+ * 1. First check explicit triggers (high confidence match)
+ * 2. Fallback to HiveDB BM25 scoring for semantic matching
+ *
+ * @param userMessage - The raw user message
+ * @returns Array of selected skills
+ *
+ * ALGORITHM:
+ * 1. If conversational → return []
+ * 2. Check explicit triggers from all enabled skills
+ * 3. If trigger match found → return matching skill immediately
+ * 4. Query the HiveDB capability index with the raw message
+ * 5. Keep hits scoring at least RELEVANCE_RATIO of the top hit
+ * 6. Hydrate skill details from HiveDB and return top MAX_SKILLS_PER_TURN
+ */
+export async function selectSkills(userMessage: string): Promise<SkillDescriptor[]> {
+    const startTime = performance.now()
+
+    log.debug(`[skill-selector] Processing user message: "${userMessage.substring(0, 100)}"`)
+
+    // Step 1: Check if conversational
+    if (isConversational(userMessage)) {
+        log.debug(`[skill-selector] Conversational message, returning empty array`)
+        return []
+    }
+
+    // Step 2: Check explicit triggers first (high priority)
+    const skillsCol = await col<SkillDoc>("skills")
+    const calendarOperation = isCalendarOperation(userMessage)
+    const allSkills = (await skillsCol.scan({}))
+        .filter(e => e.doc.active && !(calendarOperation && e.doc.category === "cron"))
+        .map(e => toSkillDescriptor(e.doc))
+
+    // Check trigger match - if found, return immediately with high confidence
+    for (const skill of allSkills) {
+        if (skill.triggers && matchTriggers(userMessage, skill.triggers)) {
+            log.info(`[skill-selector] Trigger match found: ${skill.name}`)
+            return [skill]
+        }
+    }
+
+    // Step 3: Semantic matching via the HiveDB capability index
+    let hits
+    try {
+        hits = await searchCapabilities(userMessage, { types: ["skill"], k: 20 })
+    } catch (err) {
+        log.error(`[skill-selector] Capability search failed:`, err)
+        return []
+    }
+
+    if (hits.length === 0) {
+        log.debug(`[skill-selector] No matches, returning empty array`)
+        return []
+    }
+
+    // Log raw scores for debugging
+    log.info(`[skill-selector] Raw scores: ${hits.slice(0, 10).map(h => `id=${h.rawId}, score=${h.score.toFixed(2)}`).join(", ")}`)
+
+    // Step 4: Keep only hits close enough to the best match
+    const relevantHits = applyRelativeCutoff(hits, RELEVANCE_RATIO)
+    if (relevantHits.length === 0) {
+        log.debug(`[skill-selector] All results below ratio cutoff, returning empty`)
+        return []
+    }
+
+    // Step 5: Hydrate full skill details from HiveDB (already loaded above)
+    const skillMap = new Map(allSkills.map(s => [s.id, s]))
+    const result: SkillDescriptor[] = []
+
+    for (const hit of relevantHits) {
+        const skill = skillMap.get(hit.rawId)
+        if (skill) {
+            result.push(skill)
+            if (result.length === MAX_SKILLS_PER_TURN) break
+        }
+    }
+
+    const timing = performance.now() - startTime
+
+    if (result.length > 0) {
+        log.info(`[skill-selector] Selected ${result.length} skills in ${timing.toFixed(2)}ms:`,
+            result.map(s => ({ name: s.name, category: s.category })))
+    } else {
+        log.debug(`[skill-selector] No skills selected, returning empty array in ${timing.toFixed(2)}ms`)
+    }
+
+    return result
+}
+
+// ─── Minimal Skills Loader ───────────────────────────────────────────────────
+
+/**
+ * Load the skills that ride along with the minimal tool loadout.
+ *
+ * Derived, not listed: a skill qualifies when every tool it documents is in
+ * MINIMAL_TOOLS (see minimal-loadout.ts). Any other skill reaches the agent
+ * through discovery, which injects it together with the tools it teaches.
+ */
+export async function getMinimalSkills(): Promise<SkillDescriptor[]> {
+    try {
+        const skillsCol = await col<SkillDoc>("skills")
+        const skills = (await skillsCol.scan({}))
+            .filter(e => e.doc.active && isMinimalSkill(e.doc.tools))
+            .map(e => toSkillDescriptor(e.doc))
+
+        log.info(`[skill-selector] Loaded ${skills.length} minimal skills: ${skills.map(s => s.name).join(", ")}`)
+        return skills
+    } catch (err) {
+        log.error(`[skill-selector] Failed to load minimal skills:`, err)
+        return []
+    }
+}
+
+// ─── Sync Skills to HiveDB ───────────────────────────────────────────────────
+
+/**
+ * Sync all enabled skills from database to the HiveDB capability index.
+ * Should be called on initialization from gateway/initializer.ts.
+ * Replaces all `type=skill` documents atomically (delete-by-filter + one
+ * batch commit).
+ */
+export async function syncSkillsToIndex(): Promise<void> {
+    try {
+        // Step 1: Get all enabled skills from database
+        const skillsCol = await col<SkillDoc>("skills")
+        const dbSkills = (await skillsCol.scan({})).map(e => e.doc).filter(s => s.active)
+
+        if (dbSkills.length === 0) {
+            log.debug(`[skill-selector] No skills found in DB to sync`)
+        }
+
+        // Step 2: Replace all skill documents in the HiveDB capability index.
+        // name (boost 4.0) = skill name; tags (3.0) = triggers + category +
+        // tools (the strongest semantic signals); body (2.0) = description +
+        // full skill body.
+        const docs: CapabilityDoc[] = dbSkills.map(skill => ({
+            type: "skill" as const,
+            rawId: skill.id,
+            name: skill.name,
+            tags: [skill.triggers, skill.category, skill.tools].filter(Boolean).join(" "),
+            body: [skill.description, skill.body].filter(Boolean).join("\n"),
+        }))
+
+        await replaceCapabilityDocs("skill", docs)
+
+        log.info(`[skill-selector] Sync complete: ${dbSkills.length} skills indexed in HiveDB`)
+
+    } catch (err) {
+        log.error(`[skill-selector] Skill index sync failed:`, err)
+        throw err // Re-throw to inform initializer
+    }
+}
+// ─── Debug/Test Helpers ─────────────────────────────────────────────────────
+
+/**
+ * Get all enabled skills from database (for debugging/testing)
+ */
+export async function getAllSkillsFromDB(): Promise<SkillDescriptor[]> {
+    try {
+        const skillsCol = await col<SkillDoc>("skills")
+        return (await skillsCol.scan({})).filter(e => e.doc.active).map(e => toSkillDescriptor(e.doc))
+    } catch (err) {
+        log.error(`[skill-selector] Failed to fetch skills:`, err)
+        return []
+    }
+}
+
+/**
+ * Get skill by name
+ */
+export async function getSkillByName(name: string): Promise<SkillDescriptor | undefined> {
+    try {
+        const skillsCol = await col<SkillDoc>("skills")
+        const entry = await skillsCol.get(name)
+        return entry && entry.doc.active ? toSkillDescriptor(entry.doc) : undefined
+    } catch (err) {
+        log.error(`[skill-selector] Failed to fetch skill by name:`, err)
+        return undefined
+    }
+}
+
+/**
+ * Get skills by category
+ */
+export async function getSkillsByCategory(category: string): Promise<SkillDescriptor[]> {
+    try {
+        const skillsCol = await col<SkillDoc>("skills")
+        const entries = await skillsCol.findBy("category", category)
+        return entries.filter(e => e.doc.active).map(e => toSkillDescriptor(e.doc))
+    } catch (err) {
+        log.error(`[skill-selector] Failed to fetch skills by category:`, err)
+        return []
+    }
+}

package/packages/core/src/agent/stuck-loop.ts

@@ -0,0 +1,209 @@
+import type { Config } from "../config/loader.ts";
+import { logger } from "../utils/logger.ts";
+import { hashObject } from "../utils/crypto.ts";
+
+interface ToolCallRecord {
+  toolName: string;
+  argsHash: string;
+  errorMessage?: string;
+  timestamp: number;
+}
+
+interface ProgressRecord {
+  hash: string;
+  timestamp: number;
+}
+
+export interface StuckLoopState {
+  detected: boolean;
+  toolName: string;
+  count: number;
+  lastError?: string;
+  kind: "loop" | "stall";
+}
+
+export class StuckLoopDetector {
+  private log = logger.child("stuck-loop");
+  private history: Map<string, ToolCallRecord[]> = new Map();
+  private progressHistory: Map<string, ProgressRecord[]> = new Map();
+  private readonly maxHistoryPerSession = 50;
+  private readonly triggerThreshold = 3;
+  private readonly progressThreshold = 3;
+
+  constructor(_config: Config) {}
+
+  recordToolCall(
+    sessionId: string,
+    toolName: string,
+    args: Record<string, unknown>,
+    error?: string
+  ): void {
+    let sessionHistory = this.history.get(sessionId);
+    if (!sessionHistory) {
+      sessionHistory = [];
+      this.history.set(sessionId, sessionHistory);
+    }
+
+    const record: ToolCallRecord = {
+      toolName,
+      argsHash: hashObject(args),
+      errorMessage: error,
+      timestamp: Date.now(),
+    };
+
+    sessionHistory.push(record);
+
+    if (sessionHistory.length > this.maxHistoryPerSession) {
+      sessionHistory.shift();
+    }
+
+    this.log.debug(`Recorded tool call: ${toolName} for session ${sessionId}`);
+  }
+
+  check(sessionId: string): StuckLoopState {
+    const sessionHistory = this.history.get(sessionId) ?? [];
+
+    if (sessionHistory.length < this.triggerThreshold) {
+      return { detected: false, toolName: "", count: 0, kind: "loop" };
+    }
+
+    const recent = sessionHistory.slice(-10);
+    const counts = new Map<string, { count: number; error?: string }>();
+
+    for (const record of recent) {
+      const key = `${record.toolName}:${record.argsHash}`;
+      const existing = counts.get(key);
+
+      if (existing) {
+        existing.count++;
+        if (record.errorMessage) {
+          existing.error = record.errorMessage;
+        }
+      } else {
+        counts.set(key, { count: 1, error: record.errorMessage });
+      }
+    }
+
+    for (const [key, data] of counts) {
+      if (data.count >= this.triggerThreshold && data.error) {
+        const toolName = key.split(":")[0] ?? "unknown";
+
+        this.log.warn(`Stuck loop detected: ${toolName} called ${data.count} times with same args and error`);
+
+        return {
+          detected: true,
+          toolName,
+          count: data.count,
+          lastError: data.error,
+          kind: "loop",
+        };
+      }
+    }
+
+    return { detected: false, toolName: "", count: 0, kind: "loop" };
+  }
+
+  /**
+   * Record a progress snapshot hash. Use this to detect when the agent is not
+   * advancing even though no tool errors are occurring.
+   */
+  recordProgress(sessionId: string, progressHash: string): void {
+    let sessionProgress = this.progressHistory.get(sessionId);
+    if (!sessionProgress) {
+      sessionProgress = [];
+      this.progressHistory.set(sessionId, sessionProgress);
+    }
+
+    sessionProgress.push({ hash: progressHash, timestamp: Date.now() });
+
+    if (sessionProgress.length > this.maxHistoryPerSession) {
+      sessionProgress.shift();
+    }
+
+    this.log.debug(`Recorded progress hash for session ${sessionId}: ${progressHash}`);
+  }
+
+  /**
+   * Detect lack of progress when the same hash appears repeatedly.
+   */
+  checkProgress(sessionId: string, threshold?: number): StuckLoopState {
+    const sessionProgress = this.progressHistory.get(sessionId) ?? [];
+    const required = threshold ?? this.progressThreshold;
+
+    if (sessionProgress.length < required) {
+      return { detected: false, toolName: "", count: 0, kind: "stall" };
+    }
+
+    const recent = sessionProgress.slice(-required);
+    const firstHash = recent[0]?.hash;
+    const allSame = recent.every((r) => r.hash === firstHash);
+
+    if (allSame && firstHash) {
+      this.log.warn(`Stall detected: no progress for ${required} checks (hash ${firstHash})`);
+      return {
+        detected: true,
+        toolName: "NO_PROGRESS",
+        count: required,
+        kind: "stall",
+      };
+    }
+
+    return { detected: false, toolName: "", count: 0, kind: "stall" };
+  }
+
+  clear(sessionId: string): void {
+    this.history.delete(sessionId);
+    this.progressHistory.delete(sessionId);
+    this.log.debug(`Cleared stuck loop history for session ${sessionId}`);
+  }
+
+  prune(maxAgeMs: number = 30 * 60 * 1000): number {
+    const now = Date.now();
+    let pruned = 0;
+
+    for (const [sessionId, history] of this.history) {
+      const filtered = history.filter((r) => now - r.timestamp < maxAgeMs);
+
+      if (filtered.length === 0) {
+        this.history.delete(sessionId);
+        pruned++;
+      } else if (filtered.length !== history.length) {
+        this.history.set(sessionId, filtered);
+      }
+    }
+
+    for (const [sessionId, history] of this.progressHistory) {
+      const filtered = history.filter((r) => now - r.timestamp < maxAgeMs);
+
+      if (filtered.length === 0) {
+        this.progressHistory.delete(sessionId);
+        pruned++;
+      } else if (filtered.length !== history.length) {
+        this.progressHistory.set(sessionId, filtered);
+      }
+    }
+
+    return pruned;
+  }
+}
+
+export function createStuckLoopDetector(config: Config): StuckLoopDetector {
+  return new StuckLoopDetector(config);
+}
+
+/**
+ * Build a human-readable intervention message for the model.
+ */
+export function getInterventionMessage(state: StuckLoopState): string {
+  if (!state.detected) return "";
+
+  if (state.kind === "stall") {
+    return `WARNING: Has estado sin avanzar durante ${state.count} ciclos consecutivos. Revisa tu plan, intenta una herramienta diferente o pide aclaración al usuario en lugar de repetir la misma secuencia.`;
+  }
+
+  if (state.count >= 4) {
+    return `CRITICAL: Has llamado ${state.toolName} ${state.count} veces con los mismos argumentos y sigue fallando con: "${state.lastError}". El usuario será notificado. Debes cambiar completamente de estrategia o pedir ayuda.`;
+  }
+
+  return `WARNING: Has llamado ${state.toolName} ${state.count} veces con los mismos argumentos y sigue fallando. Debes probar un enfoque diferente en lugar de repetir la misma acción.`;
+}